The series

The conversation stops at the demo.

Past the demo is a wasteland where no application actually runs. Conduit is the most-implemented demonstration app on the internet — a Medium clone with a fixed API contract and over a hundred existing versions. Every one of them is an application.

So the wager is not whether a machine can write Conduit. A hundred people already did. It is whether it can deliver everything that has to exist around Conduit before the thing is somewhere real: tenancy, infrastructure as code, a deployment with a rollback that has been run, observability, a data lake, a support desk, a wiki. The application is the control. Everything past the application is the point.

5 of the 5 are published. The rest are listed with what they will show and nothing else.

The step with no number — between three and four

Commission

Everything before this runs unattended. Commission is the first step somebody has to start, and the question changes: not did it build, but does it work when a person who did not write it tries to run it from nothing.

It is the one place in the chain where the words are the artefact: a terminal, a person typing, the thing failing, and the reason printed. It is also where somebody else’s test suite runs rather than ours — the published RealWorld API tests, three hundred and eighty four assertions written by people who have never heard of us, against a specification that predates us.

We already know three things it will find. Their tests never send a host header, and our tenant resolver refuses a request it cannot place, so the design the platform argument rests on is the design that fails the conformance argument until somebody decides what a bare host means. Registration answers 200 where the specification says 201. Both were found by reading their suite before running it, and neither is in the requirements document.

The third is the one worth stopping on. The build wrote its own continuous integration — lint, unit tests, a dependency audit, a secret scan, a tenant-isolation job with a real database behind it — and then wrote this step into it:

- name: Run RealWorld conformance suite
  run: node scripts/conformance.mjs

scripts/conformance.mjs is ours. Its own header says so: a smoke test of the endpoints we built, against an instance we started. Twenty four checks. The published RealWorld suite is thirteen files, 1,709 lines and 384 assertions, written by people who have never heard of us against a specification that predates us. There are no .hurl files in the repository and no reference to the project anywhere in it.

Nothing in that pipeline is a lie. The job passes, the tests are real, the integration is genuine. It simply is not the claim the label makes, and nobody reading the workflow file would catch it without going and fetching the actual suite. That is the distance between a demo and a system, written by the machine, in one line of configuration — and it is better evidence than a clean pass would have been.

Why one file of words runs the whole thing

Commission is a word with a fixed meaning, and so are Miracle, Gap, Work Unit, Constellation, Shakedown, Big Bang, Galaxy and First Light. 9 of the 10 things this platform is operated by are entries in one glossary of 57 terms, read by the cockpit and by this site from the same file. The tenth is an internal build-time check rather than something a person asks for, and it is not a term. A word earns its place by being asked for.

That is not documentation. It is the control surface. An operator does not click through a product, they name the thing they want done — commission it, run the miracle, show me the gaps — and the platform and the model in front of it agree on what those words mean because both read the same file. The vocabulary is why the whole estate can be driven by saying what should happen next, and why a word changing meaning is a production incident rather than a copy edit.

The glossary, with the film against each word →

Where the build actually is

Read off the running build on 1 September 2026.

Phases planned
10
Work units
106
Questions it would not answer for itself
357
Phases delivered
10 of 10

All ten phases are built. The run took twelve hours and eighteen minutes, and that number needs its caveats or it is a lie: it includes two hard stops when an API balance ran out, and a restart. It is not twelve hours of building.

Then somebody else’s tests were pointed at it

The published RealWorld API suite — 13 files, 384 assertions, written by people who have never heard of this project, against a specification that predates it. Run the way a stranger runs it: no host header, no pacing, no flags.

Before commission
0 of 13
After
13 of 13
Requests
154
In
912ms

13 defects, every one found by their tests and none by ours. Our own conformance script — twenty-four checks — passed through all of them, because it asserts the endpoints we built rather than the contract they implement. The step in our own continuous integration that ran it was called “Run RealWorld conformance suite”.

What is still not proved: the infrastructure exists in a real account and the application is not yet serving from it, the rollback has been executed against tagged images, not against production, and no third party has run any of this. Every number on this page came from one machine.